Object detection and classification methods, supervised and unsupervised, are widely used in Multi/ hyper bands image processing whose spectrum of each spectral pixel is considered to be discrete (4 to 10 bands) for multi band images, or continuous (over hundreds of bands) for hyperband images. However, in some application frameworks such as analysis of galaxies in astronomy or identification of biological cells, etc., the lack of the learning base requires the use of unsupervised approaches that will remain effective. The basic criterion of these methods is looking for pixels having different spectral characteristics from the background, these pixels are named "anomalies". These are "outliers that deviate from other observations to the point of arousing suspicion of having been generated by other mechanism" (Hawkins, 1980). In this report, we present the "Anomalous Component pPursuit (ACP)", an unsupervised statistical method involves the detection and discrimination of the rare objects in a hyperspectral image. The cited method combines two approaches: hypothesis testing (HT) with a constant false alarm rate (CFAR) and Projection Pursuit (PP) algorithm based on the independant component analysis (ICA) with the kurtosis maximization criterion. This method is applied to synthetic hyperspectral images including extended objects or extended truncated objects at several levels of noise in order to evaluate the performance and robustness according to the noise level, and the different categories of objects.
展开▼
